Sponsored by

Welcome, humans.

Fable 5 is back. That’s the publicly available “Mythos-class” (AI industry for “big ole baddie”) model from Anthropic, who is largely considered the leading AI model developer right now (but to quote Elf, “you never can tell, kid”).

As expected, the internet immediately handled the news with the calm, sober restraint of people who had just been handed a Ferrari with a speed limiter and a parent in the passenger seat. heavy on the sarcasm.

That said… there is that whole problem of Fable 5, which is supposedly the best coding agent no longer being held back by the government, actually falling back to Opus 4.8 (the previous version) for tasks including, y’know, coding. Which led to this little beauty:

Another Anthropic thread had the purest version of the mood: great, Fable is back for everything except coding... so, uh, what exactly are we using it for again? GPT 5.6 wen?

More seriously: everyone wants the magic model, but may very well end up disappointed due to the weeks of hype hyping it up. Reviews so far are pretty glowing. IMO, it’s impressive in how far you can push it and it will actually work… try to do something REALLY hard with it this holiday weekend; we’ll break it all down below!

Here’s what happened in AI today:

• 🙀 Anthropic relaunched Fable 5 after a government-driven shutdown forced new safeguards.

• 📰 Meta is very serious about launching a cloud business.

• 📰 Together AI raised $800M to scale open-model infrastructure (woohoo!).

• 🍪 xAI Voice Agent Builder launched no-code Grok Voice agents for support, sales, and scheduling calls.

• 🎓 Map the fog before starting your next big AI project.

...and a whole lot more that you can read about here.

Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us! 

P.S: We’re going live later today @ 10am PT to talk AI drama and what you should do about it all. Join us here.

Fable 5 Is Back, and Everyone’s Already Stress-Testing It

Claude’s most chaotic model week has reached the “okay, try it again” phase.

After launching Fable 5 on June 9, pulling it on June 12, and restoring access on July 1, Claude posted the simplest possible update: “Fable 5 is back.”

Here’s what happened:

• Anthropic said Fable 5 is now restored after export controls were lifted.

• The relaunch adds a new cybersecurity classifier, a filter that catches risky requests before Fable answers.

• If a request gets flagged, users are routed to Opus 4.8 instead.

• Paid users can try Fable 5 through July 7 for up to 50% of weekly usage limits, according to Claude’s help center.

How to try it:

1. Open Claude on web, desktop, mobile, Cowork, or Claude Code.

2. Pick “Fable 5” from the model selector.

3. In Claude Code, make sure you’re on version 2.1.170 or later.

4. If a normal coding request gets flagged, use /feedback in Claude Code or the thumbs buttons in Claude.ai.

The crowd reacts: The early reaction split is the interesting part. Cursor said Fable 5 is available again and leads every model on CursorBench, though it is also the most expensive per task. Theo said Fable had not rerouted him on real coding work and that the concerns were “massively overblown.” Aniket Panjwani had the most practical advice: use the one-week window for planning, hard problems, and project reviews, then hand implementation to cheaper models.

Then there is the skeptic camp. Steve Krouse said he used Fable nonstop, returned to Opus, and barely noticed the difference. Ethan Mollick had the opposite read: Fable changes the job from steering every step to commissioning a finished outcome.

Why this matters: Fable 5’s relaunch is a live test of whether frontier models can be both powerful and equitably distributed. The model is back. Now users get to find out whether the guardrails feel like seatbelts or speed bumps. So far, it seems like the new of Fable’s coding death has been overblown. You now have 7 days to use it before it switches to pay to play.

Our take: The smartest move this week is use Fable as much as possible where the work is ambiguous, long, or judgment-heavy, then send its conclusions and planning and design work to cheaper models like Opus (which is hilariously ironic to call cheaper but relatively speaking), GLM 5.2, Kimi 2.7, or OpenAi’s Codex models to implement.

Oh, and if you’ve never vibe-coded an app before and want to give it a shot? Treat yourself to the $200 Max plan of Claude this weekend and have it work on something fairly ambitious for you (planning, scaffolding, etc). You’ll be shocked how much it can handle for ya. Then you can bring in smaller models a week later to add new features. This is basically your only chance to access this level of coding intelligence this cheaply; at least for a long time. Worth it to start something big IMO!

FROM OUR PARTNERS

Scale your customer experience across all channels

Most teams scale customer service by hiring, but with ElevenAgents, real-time voice and chat agents handle high volumes of interactions and resolve them faster.

When routine requests come in, they get answered in the moment, in 70+ languages. However, the complex ones route to humans with full context, so your team focuses only on the conversations that need judgment. The experience holds up because the agents sound human, not robotic. And with Expressive Mode, they read tone and adapt - de-escalating, reassuring, and guiding each conversation to a clear resolution.

You can train every agent on your knowledge base, SOPs, and policies, deploy across voice, chat, and WhatsApp, then keep improving with real-time CSAT tracking, A/B testing, and Guardrails.

The outcome is the one leaders care about: higher resolution rates, higher CSAT, and support that scales with demand instead of headcount. Pricing is transparent and flat at $0.08 per minute, with no add-ons to work around.

Get started free

🎓 AI Skill of the Day: Map the Fog Before You Build

Big projects fail fastest when you ask AI to “start building” before you know what is still unknown.

Matt Pocock shared a planning skill he is calling /decision-mapping or /pathfinder. The move is to make the model map the “fog of war” first: the decisions, research gaps, and unknowns that could derail the project later.

Use it before a greenfield project, a messy client ask, or any workflow where the next step feels obvious but undercooked. Ask AI to separate what is fixed from what needs research, a prototype, expert input, or parallel work.

I’m starting a large project and want to map the fog before building.

Project goal:
[describe the outcome]

Known context:
[paste requirements, constraints, stakeholders, links, or notes]

Act like a senior project planner. Create:
1. The decisions that are already fixed.
2. The decision frontiers: choices that still shape the project.
3. The fog-of-war questions: unknowns that could change the plan.
4. For each unknown, label the best next move: research, prototype, ask an expert, user test, or delegate.
5. A parallel work plan with 3-5 tracks that different people or agents could run at the same time.
6. The next three actions I should take today.

Have a specific skill you want to learn? Request it here. 

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *Your AI roadmap needs a test course. The Dell Pro Max with GB10 helps teams experiment before making bigger bets. Want to run local AI? Check it.

2. xAI Voice Agent Builder lets you build no-code Grok Voice agents for support, sales, scheduling, and workflow handoffs, $0.05/minute.

3. Acti turns your phone keyboard into an agent that can fetch links, docs, schedules, profiles, and meeting links without leaving the text field, no pricing details.

4. Cognition launched Devin Security Swarm, which scans codebases, tests exploitability in sandboxes, and opens remediation PRs.

5. Hugging Face and Cerebras show how to build a low-latency open voice stack for robots and voice apps.

6. Adam CAD Copilot helps you edit Onshape and Autodesk Fusion parts with prompts while keeping the CAD model parametric and editable, free tier included.

New interview: Can AI Agents Learn From Expert Corrections?

OpenAI and Thrive Holdings built Tax AI, a Codex-powered agent that helps prepare complex tax returns while preserving evidence for accountant review.

The interesting part is the improvement loop: practitioner corrections become structured signals, Codex investigates traces and evals, and engineers review scoped fixes before anything ships.

In our latest episode, Corey and Grant talk with OpenAI’s John de Wasseige and Arthur Fernandes Araujo about what Tax AI actually does, why tax prep is such a brutal test case for agents, and why expert feedback loops may matter more than flashy demos.

Watch / Listen: YouTube | Spotify | Apple Podcasts

📰 Around the Horn

• Meta reportedly explored selling excess AI compute through a cloud business as investors looked for infrastructure returns.

• Together AI raised $800M at an $8.3B valuation to expand open-model infrastructure (someone needs to give Fireworks a new round, they rule).

• Google shipped a July 1 agent stack across Genkit, ADK 2.0, and cloud-local ML work in VS Code.

• GitHub added Copilot CLI auto model selection, which routes tasks based on reliability and cost signals.

• Hugging Face highlighted what are called “metacognition adapters” that can estimate when a model may be wrong without retraining the base model.

• Senior SWE-Bench launched as an open-source benchmark for testing coding agents on vague, long-horizon senior engineering tasks.

• Factory AI introduced Droid Shield 2.0, a learned secret-detection system for safer autonomous engineering agents.

FROM OUR PARTNERS

Get IT tickets solved faster

Thread turns IT service desks into AI-powered engines, automating triage, surfacing context, and helping teams resolve tickets faster.

🧩 Thursday Trivia

You know the drill. One is AI, and one is real. Which is which? Vote in the poll below!

A

B

A Cat’s Commentary

Trivia answer: B is AI (full video, very cool, cost $6K to make!) and A is real.

P.P.S: Love the newsletter, but only want to get it once per week? Do not unsubscribe; update your preferences here.